Change Management for Agentic AI: Preparing Your Operating Model

AI-enabled agents are entering the enterprise irrespective of whether anyone has determined what they mean for how the organisation actually works. Gartner forecasts that 40% of enterprise applications will be integrated with task-specific AI agents by 2026, up from fewer than 5% in 2025.¹ The same firm expects more than 40% of agentic AI projects to be cancelled by the end of 2027, attributing those failures to escalating costs, unclear business value and inadequate risk controls rather than to any shortcoming in the models themselves.² Read together, the two forecasts describe a single organisation: one that has deployed agents considerably faster than it has redesigned the operating model around them.

The gap is measurable. Kyndryl’s 2026 People Readiness Report, a survey of 1,100 senior business and technology leaders across eight countries, found that 79% agree the speed of AI will outpace their organisation’s workforce, governance and operating models.³ In the context of agentic AI, that mismatch ceases to be a forecast and becomes an operational problem.

Why agentic AI presents a different change problem

Generative AI, for all the attention it attracted, largely left the operating model intact. A person still performed the work and the tool made them faster. Agentic AI removes that assumption. An agent does not assist with a task; it assumes the task, executes a sequence of steps and exercises judgement along the way. The organisational implication moves from people working differently to work no longer being performed by a person at all, with someone now required to govern what the agent does in their place.

This is an operating-model change in the strict sense. It alters who performs the work, who decides, who is accountable when a decision proves wrong, and how exceptions are escalated. No deployment plan will answer those questions, because they are not technical questions.

The questions to settle before the agents arrive

The organisations we see struggling have almost invariably skipped these. Those that succeed have written their answers down:

• Decision rights. What may an agent decide on its own? What must it recommend for human approval? What must it never touch at all? Ambiguity on this point is where both cost and risk tend to escape.
• Accountability. When an agent makes a poor call, who owns the outcome: the process owner, the team or the vendor? Attributing the decision to the system is not a governance position.
• Exception handling. How quickly is a person brought back into the process, and do employees genuinely believe they are authorised to override the agent? Many do not, and often say so only after something has gone wrong.
• Role redesign. What is the human contribution now? Oversight, judgement, exception handling and relationship work have to be designed and named as substantive roles, not left for people to infer while they speculate about their futures.
• Assurance. How will the organisation establish, six months after deployment, that the agent is still doing what was intended?

Why the cancellations happen

The three reasons Gartner gives for cancellation, being cost, unclear value and weak risk controls, are governance failures rather than engineering ones.² There is a procurement dimension as well. Gartner notes the prevalence of “agent washing”, the rebranding of existing assistants, chatbots and robotic process automation as agentic, and estimates that only around 130 of the thousands of vendors claiming agentic capability are genuine.² An organisation without a clear view of the operating-model change it wants is poorly placed to judge which of those vendors it actually needs.

What the pacesetters do differently

Kyndryl’s research identifies a group comprising 9% of organisations, which it terms “pacesetters”, that do three things the remainder do not. They redesign roles around AI rather than adding AI capability to unchanged job structures. They run deliberate change management, so that the workforce understands the new operating model and guardrails are in place. And they build workforce readiness rather than assuming it. Those organisations were 1.5 times more likely to achieve AI-related revenue growth and 1.6 times more likely to report stronger innovation in products and services.³ All three behaviours are change implementation disciplines rather than technology tasks.

Preparing your operating model

This is the work CCG exists to do, and with agentic AI it has to happen before deployment rather than after:

• Set the governance and decision rights first. CCG Advisory works with Boards and executives on the oversight model, the guardrails and the accountability questions that determine whether an agentic programme is governable at all.
• Redesign the roles and the workflow. Through CCG Consult we translate an agentic ambition into changed processes, escalation paths and clearly defined human roles, which is the operating-model shift itself.
• Build the judgement the new roles demand. Supervising an agent is a different skill from performing the task. CCG Learn develops the confidence and capability people need to oversee, question and override these systems well.
• Measure whether it is landing. CCG Analytics tracks readiness, adoption and sentiment, so that resistance and risk become visible while there is still time to act, rather than at the cancellation review.

Agentic AI will reward the organisations that treat it as an operating-model decision rather than a purchase. The technology will continue to improve without intervention. Decision rights, accountability and the human roles around the agents will not; those have to be designed, and they have to be led.

If you are preparing your operating model for agentic AI, we would welcome the conversation. Get in touch with CCG.